Serge Tkach published a piece this week on how service vendors are being rebuilt around AI. The pattern he describes is a roll-up. Private-equity money is buying fragmented outsourcing firms, the help desks and claims teams and back-office shops that most large companies run through a supplier, and re-platforming them on agents. When a firm is rebuilt this way, the commercial pitch changes with the plumbing. You stop paying for seats or hours. You pay for the result. On paper that reads as the buyer-friendly successor to the seat licence, and that is how it is sold.
One fact rarely comes up in the sales meeting: the firm you contract with is often not a single company. It is several acquired shops, support desks and contact centres and finance-ops teams, each with its own data practices and security posture, running on a shared AI layer that was bolted on fast. What binds them commercially is the new price. Tkach’s own example was a support contract billed, in his words, “per resolved case, not per seat”, the kind of pricing procurement approves without a second meeting. The large integrators are doing the same, pulling labour out of their own delivery and rewriting deals around a share of savings instead of time and materials.
His example is a support contract he ran in the public sector. The vendor reported a high resolution rate and billed on it. When his team pulled the reopen data, they found auto-closed tickets logged as resolved. Users had given up raising issues because nothing came of them. The agent closed the ticket; the problem stayed open. The vendor billed on a number only the vendor could produce.
Tkach draws the conclusion plainly. “The ‘pay for outcomes’ deal is a risk transfer unless you can govern it.” Outcome pricing, he writes, “hands the advantage to whoever can measure the outcome, and in most shops, that isn’t the buyer.” The gap between the number on the contract and what your users actually live with, he argues, “is the whole game now.”
It landed in the same week the trade press was arguing about whether SaaS is dead. ZDNET put the question to Workday; ISG’s Robert Kugel argued the “SaaS-pocalypse” won’t happen. That debate is about which vendors survive the agent era, and it is the wrong thing to watch. Whether your incumbent lives or dies, the pricing model in front of you at the next renewal is changing, and the buyer’s problem was never survival. It is the meter.
Three rungs
Strip the marketing off any “outcome-based” contract and it rests on one of three things.
The top rung is a true business outcome: profit, or cost per claim settled. It lives in your books. Your vendor cannot see it.
The middle rung is a proxy: resolved conversations, deflection rate, tasks completed. It is countable, it bills monthly, and it sits inside the vendor’s system.
The bottom rung is raw consumption, priced per run or per action.
Almost no deal is written on the top rung. It rarely makes it into the signed contract, and the reasons hold on both sides. You cannot cleanly attribute a profit swing to one vendor when the market moved and three other initiatives shipped the same quarter. The vendor will not price against numbers it cannot audit, because your books are its moral hazard. And your finance team needs an invoice every month, not a P&L reconciliation every year. So the deal is sold in top-rung language and signed on middle-rung units that the vendor defines and counts. That is the risk transfer. Tkach’s resolution rate was the proxy standing in for the outcome, and the proxy was gamed.
There is an obvious objection, and it is a good one. Surely the client measures its own P&L better than any vendor could. Yes, and that is exactly why the P&L is never the thing on the invoice. The outcome you care about sits in your accounts, and no vendor bills against a number it cannot see. What you have bought, when the meter belongs to the seller, is activity pricing with better marketing. It becomes outcome pricing only when the proxy can be checked from your side.
And once money rides on a number, the vendor optimises the number: an agent paid to close tickets will close tickets, served customer or not, which is why the audit right matters more over a three-year term than the choice of metric on day one.
That is the whole test. Can the billable unit be verified from the buyer’s side of the log? The reason the auto-closed tickets got caught is that reopen rates were the client’s own data, held in the client’s own system. A well-chosen proxy is client-verifiable: a resolution that stays closed for fourteen days, counted in your ticketing platform rather than the vendor’s dashboard. A badly chosen one can only be confirmed by the party sending the bill.
The software vendors are next
None of this stays in the outsourcing corner. IDC’s forecast is that “by 2028, pure seat-based pricing will be obsolete, with 70% of software vendors refactoring their pricing strategies around new value metrics, such as consumption, outcomes, or organisational capability.” Read that list of value metrics again. Most of the 70% is consumption and proxies, not profit-share. Bain, cited in the same reporting, expects buyers to demand outcome-based over access-based pricing as agents take over discrete tasks, and names Intercom and Salesforce as already moving. The seat is going. What replaces it, for most buyers, is a meter someone else reads.
Falling model prices will not rescue the budget either. Four token price cuts landed this month, three of them inside eight days, and enterprise agent bills still went up. Janakiram MSV set out the arithmetic: total spend is task volume, times attempts per task, times tokens per attempt, times price per token, plus tool and infrastructure cost. Only the fourth term is falling. Cheaper inputs do not give the buyer control of the bill, which is the point issue 52 made about inference costs, now one level up.
Five questions before you sign
If a renewal or a new agent contract puts consumption or outcome terms in front of you this half, the diligence is short and specific.
- What is the billable unit, defined in writing, with its edge cases named?
- Who produces the count, and from which log?
- Do you get the raw event log, or only the invoice?
- What is the dispute path, and who corrects an over-count once one is found?
- What does the meter read at exit: accrued units, in-flight work, and a final figure?
Issue 56 covered the contract clauses that can stop an agent programme before it starts. This is the next term in the same negotiation: the price you are charged once it runs.
For Polish leadership
For most Polish enterprises, these terms show up first in an H2 2026 renewal rather than in a new purchase. If you are supervised by the KNF, you already maintain a DORA third-party register under Article 28. How a critical AI vendor’s bill is measured, and whether you can audit that measurement, belongs in the same record as everything else you track about that supplier. Treat it as one more line in that register: the vendor, the critical function it supports, and the log you would pull to contest an invoice. The work is small at signing and close to impossible once the contract is live and the vendor holds the only copy of the count. There is a first-mover point here as well. No Polish company has published how it audits outcome-based AI pricing. The firm that writes that internal standard first will be the one the others quote.
The question for your board is narrow. On every agent contract you are about to sign or renew, who owns the meter, and can you read it?
Briefing
Nadella warns that your corrections are training your competitor. Satya Nadella described a “Reverse Information Paradox”: every time an employee corrects an AI tool, institutional know-how leaks back to the model provider. He joins Palantir’s Alex Karp in the warning. So what: the data your staff feed a vendor’s model is a board-level governance question, not a procurement footnote. Fortune
SAP buys Prior Labs in a €1bn bet on tabular AI. SAP closed its acquisition of the maker of TabPFN, a foundation model built for structured data rather than chat. So what: the AI that matters for regulated back offices reasons over tables, and the incumbents are paying up to own that layer. The Next Web
Reflection AI signs a $1bn compute deal before shipping a model. The firm committed over $1bn to Nebius for capacity through 2029, ahead of releasing an open-weight model meant to rival China’s open-source labs. So what: frontier capability is gated by compute contracts signed years in advance, which is a measure of how durable today’s supplier concentration is. Forbes
Summary
Outcome-based pricing is replacing the seat licence, sold as the fairer deal. In practice the real business outcome sits in the buyer’s books, where no vendor can bill against it, so contracts are signed on proxies the vendor defines and counts. That moves the measurement risk onto the buyer. The auto-closed tickets in Tkach’s example were caught only because the reopen data was the client’s own. The test for any consumption or outcome contract is whether the billable unit can be verified from your side of the log. For a firm supervised by the KNF, that verification belongs in the DORA third-party register, alongside everything else it tracks about a critical supplier. Read the meter before you agree to be priced on it.
Stay balanced, Krzysztof Goworek
Krzysztof Goworek is founder of Quintant — AI advisory that gets enterprises from experiment to production value.